Comparative Performance Analysis of Kernel Functions in Support Vector Machines in the Diagnosis of Pneumonia using Lung Sounds

Abstract

Pneumonia, according to the World Health Organization (WHO) is globally one of the leading causes of deaths among communicable respiratory diseases. It is caused by a viral, bacterial, or fungal infection affecting the lower respiratory tract causing the alveoli to be filled with fluid and pus. Chest X-rays in Pneumonia diagnosis is the current standard in practice today. These diagnostic systems are immobile, radiative, expensive, and need a clinical expert to interpret results and hence inaccessible to most living across different geographical locations and financial states. In recent years the early diagnosis of pneumonia using Lung Sounds has been under extensive research and approached clinically using various test methods by healthcare professionals. Lung sounds have certain challenges as adventitious sounds and acoustic noise are hard to differentiate. Another major challenge is that most developed systems are not optimized to work exclusively for lung sound diagnostic applications. In this work, we have approached to address both these issues. ICBHI Respiratory Database is used as a source in which Healthy and clinically labeled cases of Pneumonia were used for this study. Various signal processing algorithms were tested on filtering noise. Mel Frequency Cepstral Coefficients (MFCC) including the Energy parameter were extracted from each lung sound recording for the analysis. Support Vector Machine using different kernel methods was developed and performance tested. Radial Basis Function Kernel provided a good margin of separation compared to sigmoidal, polynomial, or linear kernel methods. Comparative performance analysis of various Kernel Functions on the dataset was observed including and excluding the Energy parameter in the feature coefficients. The classification performance for various kernels are as follows, Sigmoid Kernel 81.48%, Linear Kernel 93.53%, Polynomial Kernel 94.01%, and the Radial Basis Function (RBF) Kernel provided an accuracy of 97.8%. RBF excluding the Energy parameter gave a low accuracy of 89.98% in comparison. It was concluded that the Radial Basis Function performed best for the diagnosis of pneumonia from Lung sounds and the energy parameter is a significant discriminatory feature